The IsActive() method returned true even if the timer was not active,
after it completed once. This broke the state file timer, and the
state file was not saved periodically.
Turns out the lock-free code using atomics was not thread-safe. The
given callback could be invoked by GLib before the source_id attribute
was assigned. This commit changes the DeferredMonitor class to use a
Mutex to block the event loop until source_id is assigned. This bug
does not exist in the 0.19 branch because it does not use the GLib
main loop anymore.
Users now have to call Close() explicitly. This simplifies using the
class, as most users have automatic socket management already, and
Steal() had to be used often.
Add a flag that indicates whether the EventLoop is currently "busy".
As long as that flag is set, it does not need to be woken up - we can
simply add the DeferredMonitor to the list, and it will be caught by
EventLoop very soon. This eliminates nearly all of the
DeferredMonitor overhead when compared to IdleMonitor, rendering
IdleMonitor mostly obsolete.
Commit 1f11959 allowed modifying the "quit" attribute from any thread,
and thus the assertion may fail spuriously. This assertion is too
strict for the relaxed use of "quit". Let's remove it and move the
"quit" check to before the SockedMonitor::Dispatch() call.
Simply set the "quit" flag and wake up the thread. This works even if
we're inside this thread. Setting "quit" to a new value without mutex
protection is usually not safe, but good enough here.
Instead of creating a new eventfd for each DeferredMonitor instance,
reuse EventLoop's eventfd, and add a std::list to EventLoop that
manages the list of pending DeferredMonitors. This std::list is
protected by the same mutex as the "calls" list.
The bottom line is: reduced overhead because the per-instance eventfd
was eliminated, slightly added overhead due to Mutex usage (but
negligible), and we're thread-safe now.
This subsystem is now good enough to replace EventLoop::AddCall().
The DynamicFifoBuffer methods never return nullptr when the buffer is
empty or full; instead, they return an empty buffer. This bug caused
an endless loop.